融合注意力和残差反馈的肺结节检测算法研究  

LUNG NODULE DETECTION ALGORITHM INTEGRATING ATTENTION AND RESIDUAL FEEDBACK

作　　者：侯柯冰 程晓荣[1] 张铭泉[1] Hou Kebing;Cheng Xiaorong;Zhang Mingquan(Department of Computer,North China Electric Power University,Baoding 071003,Hebei,China)

机构地区：[1]华北电力大学(保定)计算机系,河北保定071003

出　　处：《计算机应用与软件》2024年第10期287-292,348,共7页Computer Applications and Software

基　　金：中央高校基本科研业务费专项资金项目(2020MS122)。

摘　　要：肺结节作为肺癌早期的表现形式,且肺结节大小不一,将其快速准确地检测出来不仅对预防及治疗肺癌有重大意义而且是一项艰巨的任务。为此,提出一种改进YOLOv3的方法。针对肺结节大小不一的问题,改进多尺度特征融合结构,由3尺度预测增加至4尺度来获取更多小尺度结节特征信息;添加注意力机制提高对有效信息的提取能力,并改进回环残差结构增强有效信息加快网络的收敛速度;选用CIoU为边界框损失函数。实验分析表明,改进后的算法mAP相较于原算法提升了0.063,FPS提升了4.8帧/s,对肺结节检测效果有所提升,与其他算法相比具有更好的准确性和实时性。As the early manifestation of lung cancer,lung nodules are of different sizes.The rapid and accurate detection of lung nodules is not only of great significance for the prevention and treatment of lung cancer,but also an arduous task.To this end,a method to improve YOLOv3 is proposed.In view of the problem of different sizes of lung nodules,the multi-scale feature fusion structure was improved,which increased from 3-scale prediction to 4-scale to obtain more small-scale lung nodule feature information.The attention mechanism was added to improve the ability to extract effective information,and the loop-back residual structure was introduced to enhance the effective information and accelerate the convergence speed of the network.CIoU was selected as the bounding box loss function.Experimental analysis shows that compared with the original algorithm,the improved algorithm s mAP has increased by 0.063,and the FPs has increased by 4.8 frames/s,and the detection effect of lung nodules is improved.Compared with other algorithms,it has better accuracy and real-time performance.

关 键 词：肺结节 目标检测 残差反馈 注意力机制 CIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]